Independent · Source-cited · No sponsored rankingsSources linked in every article
AIToolRanked
AI News · 5 min read

Claude-Aided Navier-Stokes Proof: What's Verified So Far

NYU's Tristan Buckmaster and Anthropic's Levent Alpöge claim Navier-Stokes blow-up with no outside force. What's out, what's missing and why it matters.

RA
· Founder, AIToolRanked
Abstract illustration of a swirling blue fluid vortex on a dark background, its streamlines tightening into one bright point at the center
On this page

What happened on November 20

On Friday, November 20, 2026, NYU mathematician Tristan Buckmaster and Anthropic mathematician Levent Alpöge posted a GitHub repository ↗ titled Finite-time blow-up for the Navier–Stokes equations. It holds a 79-page paper dated that day and a Lean 4 formalization. The claim is that the 3D incompressible Navier-Stokes equations, which describe how water and air flow, can "blow up" in finite time from smooth starting data, with no external force acting on the fluid.

Whether that can happen is one of the Clay Mathematics Institute's $1 million Millennium Prize problems. It is also the case OpenAI's September result did not cover. Scientific American ↗ reported the new claim the same day, and The Next Web ↗ followed on November 21.

It went up the same day Alpöge announced his Lean-only Riemann Hypothesis proof with Claude.

The image above is an original illustration made by aitoolranked, not a figure from the paper.

Why the "no force" part matters

In September, OpenAI said an internal model had proved Navier-Stokes blow-up. Its construction adds a smooth external force that stirs the fluid in just the right way. That is allowed by one of the options in the official Clay problem statement, but many mathematicians saw it as a loophole. "The most important problem is unsolved," University of Chicago mathematician Luis Silvestre told Scientific American ↗ in September. "The Clay problem is settled, but the main problem for the Navier-Stokes equations is not."

The new paper goes after that main problem: blow-up driven only by the fluid itself.

There is history here. In September, Buckmaster and Alpöge posted their own forced results one day before OpenAI announced its proof, and Buckmaster accused the company of trying to scoop them. OpenAI denied it. TNW reports that OpenAI later said its investigation found Buckmaster's Codex prompts could not have influenced its system.

What was released and what is missing

The repository is unusually frank about its gaps. From its README ↗:

PieceStatus on November 22
79-page paper (version of November 20)Released
Lean 4 library (about 2,270 files, 600,000 lines)Released
Numerical certificate the proof depends onNot released; the paper gives its hash and says it will be published
Check that the certificate meets the paper's conditions (interval arithmetic, 128-bit, Arb via FLINT)Outside the Lean proof and not in the repo
November 20 patch to the Lean filesApplied, but the patched package was not rebuilt

The Lean README also says the formalization came first: "a first draft of the paper was obtained from the Lean proof," and the paper was then rewritten. TNW reports the authors used Claude and Codex throughout, and that Anthropic models wrote the first proof without force.

Buckmaster explained the rough edges on Mastodon, as quoted by TNW: "After last time, we couldn't risk waiting, so the Navier-Stokes write-up is unpolished and heavily LLM-assisted. Fortunately, once the numerical profile is in hand, the proof follows a well-trodden computer-assisted path."

Terence Tao wrote ↗ that the paper is "technically human-readable" and the strategy "broadly clear": build an approximate self-similar solution numerically, then use rigorous perturbation theory to turn it into an exact one. He added that "more work needs to be done to create a proper, professionally written proof."

On the same day, Alpöge, Buckmaster and NYU's Matei Coiculescu released a separate paper proving blow-up for the unforced 3D Euler equations (fluids with no viscosity), also with a Lean formalization. Tao called that one a human-readable proof and listed it as the most "digested" of the three results.

How this compares with the Riemann claim

The Riemann proof is further along. On November 21 the Lean and Mathlib communities published a joint statement ↗ saying they have "full confidence" Alpöge's Riemann Hypothesis proof is correct. Nothing like that exists yet for Navier-Stokes. Here the Lean code proves the result from a certificate, and the certificate and the computer check of it are both still missing. Until they are public, nobody outside the team can confirm the last step.

What people are saying

  • "As if this was not enough, Buckmaster and Alpoge now claim Unforced Euler and Navier-Stokes" — @mathandcobb ↗, in a post with about 800 likes

  • "IIUC, OpenAI did not solve the full Navier-Stokes conjecture, but Buckmaster/Alpöge are now claiming that they have." — @quasicoh ↗

  • "Of course, neither is confirmed yet... But it's clear that AI timelines in math keep landing early." — Linas Beliūnas ↗ (@linasbeliunas)

  • "79 pages, about 600,000 lines of Lean, and no certificate." — @Sabbirbyte ↗

Reddit was not reachable from our tools for this story.

Our take

It's a serious claim, but not a verified result yet. The Lean library is real and large. The authors say plainly what is missing, and Tao, who has worked on blow-up for these equations himself, calls the strategy broadly clear. But the whole thing rests on a numerical certificate nobody outside the team has seen. Until it is released and someone else runs the interval-arithmetic check, the honest wording is "claimed," not "solved." Expect that check to be quick once the file is out, because that kind of computer-assisted step is routine.

Two labs are now racing on the same famous problem. OpenAI's September proof used a force. This one, built with Claude and Codex, says it doesn't need one. If it holds up, it settles the version mathematicians actually cared about. That matters more to the AI-for-science story than any benchmark score. It also shows why Lean matters: you can check the logic without trusting the lab.

For people choosing AI tools, the useful lesson is the workflow. Both math claims this week pair a frontier model with a formal checker that catches mistakes. For everyday coding, the closest version of that setup is an agent with good tests. Our Codex vs Claude Code comparison and Claude Code review cover those tools, and ChatGPT vs Claude vs Gemini covers the chat apps. For more AI math results, see OpenAI's GPT-6 Astra science results and DeepMind's AlphaProof Code.

Verdict: check back once the certificate is out. Until then, treat it as a strong claim with a known missing piece.

More from the news desk · About Rai Ansar